--- schema: 1 type: video title: 'The Cost of Accessibility in AI — Jemima Abu, CAIS' description: >- The same Figma implementation prompt runs twice: once without accessibility guidance and once with an accessibility skill. In Jemima Abu's small experiment, the second run costs 58% less and finishes in about a third of the time. She immediately explains why that uncontrolled comparison cannot prove that accessible development is always cheaper. What it does expose is a familiar failure mode: an agent invents a verbose custom component when a native HTML element could supply much of the required behavior. Recreating semantics, keyboard interaction and expected browser behavior adds code and testing that a better starting point can avoid. Abu connects that example to the context agents inherit. An inaccessible codebase teaches the wrong conventions, and a vague design prompt leaves too many product requirements unstated. Accessibility therefore begins with business requirements, design decisions and reusable components, then continues through implementation and QA. Her guidance includes auditing existing interfaces, packaging explicit accessibility instructions into agent skills, preferring semantic HTML and checking keyboard interaction. ARIA can help where necessary, but it does not replace correct native behavior. Waiting until the product is complete creates avoidable refactoring and regression work. Automated checks also leave gaps that require human review. The talk gives teams a concrete way to improve the instructions and patterns that shape generated interfaces while retaining responsibility for the finished experience. Its cost figures remain an illustration from one experiment; the broader argument is to make accessibility a normal requirement throughout the work rather than a repair at the end. Speaker info: - https://github.com/jemimaabu - https://www.linkedin.com/in/jemimaabu/ - https://www.jemimaabu.com/ Timestamps: 00:00 - The question of accessible AI generated interfaces 1:14 - An uncontrolled accessibility skill experiment 3:42 - Agents inherit the codebase's conventions 4:31 - Native HTML instead of verbose custom components 5:45 - Accessibility from business requirements through QA 7:27 - Audits, reusable patterns and agent skills canonical: 'https://liquidgold.cc/watch/ZccJsdOE54M' markdown: 'https://liquidgold.cc/watch/ZccJsdOE54M.md' llms: 'https://liquidgold.cc/watch/ZccJsdOE54M/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:18:44.449Z' language: null children: [] videoId: ZccJsdOE54M durationSeconds: 602 publishedAt: '2026-10-14T13:00:33.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 10 endSeconds: 74 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=10' heldReason: rights - ordinal: 1 text: null startSeconds: 10 endSeconds: 74 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=10' heldReason: rights - ordinal: 2 text: null startSeconds: 74 endSeconds: 86 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=74' heldReason: rights - ordinal: 3 text: null startSeconds: 86 endSeconds: 222 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=86' heldReason: rights - ordinal: 4 text: null startSeconds: 86 endSeconds: 222 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=86' heldReason: rights - ordinal: 5 text: null startSeconds: 222 endSeconds: 233 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=222' heldReason: rights - ordinal: 6 text: null startSeconds: 233 endSeconds: 277 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=233' heldReason: rights - ordinal: 7 text: null startSeconds: 277 endSeconds: 288 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=277' heldReason: rights - ordinal: 8 text: null startSeconds: 288 endSeconds: 352 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=288' heldReason: rights - ordinal: 9 text: null startSeconds: 352 endSeconds: 364 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=352' heldReason: rights - ordinal: 10 text: null startSeconds: 364 endSeconds: 448 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=364' heldReason: rights - ordinal: 11 text: null startSeconds: 364 endSeconds: 448 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=364' heldReason: rights - ordinal: 12 text: null startSeconds: 448 endSeconds: 460 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=448' heldReason: rights - ordinal: 13 text: null startSeconds: 460 endSeconds: 602 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=460' heldReason: rights - ordinal: 14 text: null startSeconds: 460 endSeconds: 602 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=460' heldReason: rights - ordinal: 15 text: null startSeconds: 460 endSeconds: 602 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=460' heldReason: rights - ordinal: 16 text: null startSeconds: 460 endSeconds: 602 watch: 'https://liquidgold.cc/watch/ZccJsdOE54M?t=460' heldReason: rights --- # The Cost of Accessibility in AI — Jemima Abu, CAIS The same Figma implementation prompt runs twice: once without accessibility guidance and once with an accessibility skill. In Jemima Abu's small experiment, the second run costs 58% less and finishes in about a third of the time. She immediately explains why that uncontrolled comparison cannot prove that accessible development is always cheaper. What it does expose is a familiar failure mode: an agent invents a verbose custom component when a native HTML element could supply much of the required behavior. Recreating semantics, keyboard interaction and expected browser behavior adds code and testing that a better starting point can avoid. Abu connects that example to the context agents inherit. An inaccessible codebase teaches the wrong conventions, and a vague design prompt leaves too many product requirements unstated. Accessibility therefore begins with business requirements, design decisions and reusable components, then continues through implementation and QA. Her guidance includes auditing existing interfaces, packaging explicit accessibility instructions into agent skills, preferring semantic HTML and checking keyboard interaction. ARIA can help where necessary, but it does not replace correct native behavior. Waiting until the product is complete creates avoidable refactoring and regression work. Automated checks also leave gaps that require human review. The talk gives teams a concrete way to improve the instructions and patterns that shape generated interfaces while retaining responsibility for the finished experience. Its cost figures remain an illustration from one experiment; the broader argument is to make accessibility a normal requirement throughout the work rather than a repair at the end. Speaker info: - https://github.com/jemimaabu - https://www.linkedin.com/in/jemimaabu/ - https://www.jemimaabu.com/ Timestamps: 00:00 - The question of accessible AI generated interfaces 1:14 - An uncontrolled accessibility skill experiment 3:42 - Agents inherit the codebase's conventions 4:31 - Native HTML instead of verbose custom components 5:45 - Accessibility from business requirements through QA 7:27 - Audits, reusable patterns and agent skills Canonical: https://liquidgold.cc/watch/ZccJsdOE54M · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:00 to 1:26) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:26 to 3:42) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (3:42 to 4:37) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (4:37 to 5:52) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (5:52 to 7:28) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (7:28 to 10:02) > Held: transcript rights do not permit this chunk's text.